Ralph Loop

Ralph Loop is a harness.

It has a "Promise" that defines a success/failure of a task, and asks the Agent to assess the Promise when it shuts down. When it did not deliver the promise, the script simply re-triggers the task.

This was later improved to @github@karpathy@autoresearch in a way that autoresearch has a numeric benchmark.

mkdir -p scripts/ralph
cp ../ralph/ralph.sh scripts/ralph/
cp ../ralph/prompt.md scripts/ralph/prompt.md
chmod +x scripts/ralph/ralph.sh